convergedmachine/Cross-D-Conv

Cross-D Conv: Cross-Dimensional Transferable Knowledge Base via Fourier Shifting Operation

This repository introduces Cross-D Conv, a novel convolutional operation designed to bridge the dimensional gap between 2D and 3D medical imaging datasets. By leveraging the Fourier domain for phase shifting, Cross-D Conv enables seamless weight transfer between 2D and 3D convolutional operations. This method addresses the challenge of multimodal data scarcity by utilizing abundant 2D data to enhance 3D model performance effectively.

@article{yavuz2024cross,
  title={Cross-D Conv: Cross-Dimensional Transferable Knowledge Base via Fourier Shifting Operation},
  author={Yavuz, Mehmet Can and Yang, Yang},
  journal={arXiv preprint arXiv:2411.02441},
  year={2024}
}

Performance Metrics

Table 1: ResNet18 Performance on Imagenet and RadImagenet

Dataset	Model	Precision (Macro)	Recall (Macro)	F1 (Macro)	Balanced Accuracy	Average Accuracy
IN1K	Regular	0.6807	0.6693	0.6657	0.6693	0.6693
	Cross-D Conv	0.6895	0.6881	0.6838	0.6881	0.6881 ↑
RIN	Regular	0.5830	0.4989	0.5252	0.4989	0.8305
	Cross-D Conv	0.5891	0.5228	0.5471	0.5228	0.8374 ↑

Table 2: Performance on Image Datasets

Dataset	Method	OrganC Mean ± Std (CT)	OrganS Mean ± Std (CT)	Brain Tumor Mean ± Std (MRI)	Brain Dataset Mean ± Std (MRI)	Breast Mean ± Std (US)	Breast Cancer Mean ± Std (US)	Average
IN1K	2D Conv	0.862 ± 0.006	0.708 ± 0.035	0.884 ± 0.011	0.305 ± 0.023	0.819 ± 0.019	0.745 ± 0.024	0.720
	Cross-D Conv	0.871 ± 0.007	0.763 ± 0.008	0.892 ± 0.010	0.308 ± 0.026	0.836 ± 0.021	0.759 ± 0.022	0.738 ↑
RIN	2D Conv	0.842 ± 0.006	0.742 ± 0.008	0.902 ± 0.010	0.268 ± 0.023	0.832 ± 0.021	0.762 ± 0.016	0.725
	Cross-D Conv	0.848 ± 0.008	0.743 ± 0.008	0.910 ± 0.013	0.283 ± 0.023	0.835 ± 0.037	0.747 ± 0.024	0.728

Table 3: Performance on Volumetric Datasets

Dataset	Method	Mosmed Mean ± Std (CT)	Lung Aden. Mean ± Std (CT)	Fracture Mean ± Std (CT)	BraTS21 Mean ± Std (MRI)	IXI Mean ± Std (MRI)	BUSV Mean ± Std (US)	Average
IN1K	ACS-Conv	0.523 ± 0.057	0.532 ± 0.034	0.456 ± 0.027	0.539 ± 0.030	0.542 ± 0.044	0.559 ± 0.079	0.525
	Cross-D Conv	0.505 ± 0.068	0.513 ± 0.071	0.469 ± 0.027	0.549 ± 0.031	0.583 ± 0.059	0.590 ± 0.064	0.535 ↑
RIN	ACS-Conv	0.547 ± 0.072	0.548 ± 0.034	0.471 ± 0.034	0.545 ± 0.041	0.555 ± 0.046	0.604 ± 0.063	0.545
	Cross-D Conv	0.557 ± 0.102	0.529 ± 0.058	0.491 ± 0.032	0.558 ± 0.044	0.559 ± 0.050	0.602 ± 0.066	0.549

---

license: mit